NVIDIA is helping launch the Open Secure AI Alliance. One of our first contributions is NVIDIA Labs Object-Oriented Agents (NOOA), and I think this one is extremely cool.
Open models matter - but open weights and data alone do not make the whole agent inspectable.
Basically: The model is one piece. The harness decides what context the model sees, what tools it can call, what state it keeps, and when it stops. That can change the results of tasks a LOT without changing the model.
NOOA makes the harness look like regular Python:
- methods = capabilities
- fields = state
- docstrings = prompts
- type annotations = contracts
So you can diff it, review it, unit test it, trace it, version it, refactor it - normal software stuff.
This is a research preview - and not a replacement for the harness you already use. The point is to put the implementation and evals somewhere people can actually inspect them, reproduce the work, disagree with it, and make it better.
The Open Secure AI Alliance is all about doing more of that across the industry: sharing practical security research - about models, harnesses, tools, evals, and other learnings - so other teams can inspect it, build on it, and make it better.
Resources: